Netflix-Style AI Can Now Forget You in 2% of the Time

Every time you tap "don't recommend this" or ask a company to delete your account, you probably assume the AI forgets you. It usually doesn't. Recommendation systems — the software behind Netflix rows, Spotify playlists and TikTok feeds — learn from your history, and that learning gets baked into the model itself. Under Europe's GDPR, which grants a legal "right to be forgotten," companies are supposed to remove that influence. The honest way to do it is to retrain the entire model from scratch without your data. That can take days and cost serious money, so it rarely happens.

A team of Italian researchers has a shortcut. Their system, called SURF, doesn't rebuild the whole thing. Instead it finds the small neighborhood of items connected to the one you want forgotten — think of finding the right drawer in a giant filing cabinet rather than emptying the whole cabinet — trains a tiny helper model on just that drawer, then subtracts the helper's suggestions from the main model's answers. The memory gets cancelled out, like noise-cancelling headphones removing a sound.

In tests across seven datasets, SURF matched the results of a full retraining while using only 2% of the computing time. On one measure of recommendation quality, it performed up to 32% better than existing methods. The researchers also published their code, meaning other companies can adopt the approach rather than reinvent it.

The catch: SURF is "approximate." It doesn't guarantee your data is completely gone — faint traces could linger in ways these tests don't catch, and regulators may eventually demand proof rather than close enough. It also still requires some computing work, just far less. Even so, it's a meaningful step toward deletion requests that actually mean something, rather than a form letter that quietly does nothing.